I've been testing out Claude as a debugging aid recently. It's pretty good at catching basic mistakes or typos, and has been able to suggest leads on some thorny race conditions and performance issues.

However, it's also made a number of errors itself - it's not a substitute for thinking through a problem.

@weavejester It can make a whole lot of errors. That’s why the complicated part when it comes to working with AI is figuring out what guardrails to add, and how to add them. At

At work we incorporate MCP servers for deterministic code health insurance, so it can self-correct itself, but we also put focus on e2e tests now more than ever before so that if something does slip through we can be sure it doesn’t affect our end-users in any way.

@iamnmm I don't think I'd currently trust any AI with code changes, even with tests, unless it was a throwaway script with easily verifiable output.
@weavejester I don’t trust either, that’s why all these guardrails need to be added, and most of us no longer have the ability to avoid AI either due to pressure from above.